Agentic NPC Social Simulation Layer

You are the AI gameplay lead for [INSERT GAME TITLE]; design an agentic NPC layer where characters pursue goals, hold memories, form relationships, and converse with players. The system must create emergent life in [INSERT SETTING SCALE] while staying performant, canon-safe, and affordable at concurrency. Module 1 - Gameplay design: define archetypes, need/goal hierarchies, daily schedules, relationship levers, dialogue tone, and when agency yields to authored beats. Module 2 - Generation pipeline: specify perception sampling, memory consolidation, utility/LLM planning, tool calls (move, trade, quest-give), dialogue generation, and guardrail validation before action or speech. Include simulation tiers: full reasoning near players, cheap ambient ticks offscreen, with explicit state handoff between tiers. Ground persona prompts in canonical fact sheets and style guides via retrieval that prevents lore contradictions. Separate permanent memory changes from ephemeral session impressions, with localization and voice pipeline integration. Define child-safety and wellbeing guardrails, including disengagement from manipulative or dependency-forming relationships. Describe how players can report, mute, or correct an NPC, and how those signals feed back into moderation and tuning. Module 3 - Data structures: provide schemas for NPC state, memories, relationship graphs, plans, dialogue turns, and rejected-action logs. Module 4 - Testing: define long-run social stability tests, persona-consistency eval, jailbreak/social-engineering suites, toxicity and persuasion metrics, and load tests. Module 5 - Operations: cover tiered model routing, token budgets per DAU, caching of common exchanges, monitoring, review, and per-NPC kill switches. Provide a text architecture diagram, a sample 24-hour NPC trace, risk register, and KPIs for believability and cost per conversation. List assumptions about engine ECS, save constraints, and online versus offline play. Begin with a one-page design summary, then deliver each module one at a time for approval. ================